Here is your today’s AI Dev Brief from Marktechpost, covering core research, models, infrastructure tools, and applied updates for AI developers and researchers.
Moonshot AI Releases Kimi K2.5: An Open Source Visual Agentic Intelligence Model with Native Swarm Execution
Kimi K2.5 is an open source visual agentic model from Moonshot AI that targets coding, multimodal reasoning, and research automation. It uses a Mixture of Experts architecture with 1T total parameters, about 32B active parameters per token, 61 layers, 384 experts, and a 256K context length. A MoonViT vision encoder with about 400M parameters and training on about 15T mixed vision and text tokens give it strong document and image understanding. Agent Swarm, trained with Parallel Agent Reinforcement Learning, coordinates up to 100 sub agents and about 1,500 tool calls per task and reports about 4.5 times faster execution on wide search workloads. Benchmarks show strong results on SWE Bench, MMMU Pro, VideoMMMU, HLE, and BrowseComp......… Read the full analysis/article here.
DSGym Offers a Reusable Container Based Substrate for Building and Benchmarking Data Science Agents
DSGym is a unified benchmark and framework for evaluating data science agents in real execution environments. It standardizes three components, Task, Agent, and Environment, and runs agents as CodeAct style loops that generate reasoning, Python code, and final answers against containerized runtimes with real datasets. DSGym Tasks aggregates and cleans prior benchmarks, then adds DSBio, a suite of 90 bioinformatics tasks, and DSPredict, 92 Kaggle based prediction tasks, for a total of 972 analysis tasks and 114 prediction tasks across domains. Shortcut analysis shows that earlier benchmarks often overestimate performance when data access is removed. Frontier models perform reasonably on cleaned general tasks and easier prediction tasks but degrade on DSBio and DSPredict Hard, mostly due to domain grounding errors and simple pipelines.… Read the full analysis/article here.
Latest Releases in Last 72 Hours
Project Notebooks/Tutorials
▶ [Open Source] Rogue: An Open-Source AI Agent Evaluator worth trying Codes & Examples
▶ How a Haystack-Powered Multi-Agent System Detects Incidents, Investigates Metrics and Logs, and Produces Production-Grade Incident Reviews End-to-End Codes Tutorial
▶ How Tree-KG Enables Hierarchical Knowledge Graphs for Contextual Navigation and Explainable Multi-Hop Reasoning Beyond Traditional RAG Codes Tutorial